mirror of
https://github.com/ether/etherpad-lite.git
synced 2026-07-17 16:47:05 +00:00
8dbd917718
* test(ci): kill Windows + Node 24 backend-test flake; capture native crashes The Backend tests suite has a ~22% silent-failure rate on Windows + Node 24 specifically (Linux 22/24/25 ✓, Windows 22/25 ✓). Two prior PRs instrumented the failure — common.ts handlers (#7663), then an unconditional diagnostics.ts (#7665) — and confirmed it's a hard kill: diagnostics.ts:23-27 documents the matrix, and every recurrence (run 25279692065, 25754938013, 25906496503) shows only `[diag +0ms] diagnostics loaded`, no beforeExit / exit / unhandledRejection / uncaughtException / signal handlers. Process dies 700–900 ms after the last passing test, in a varying spec each time. That's a native crash in V8 / libuv / the tsx loader, not anything reachable from JS. This PR ships two independent attacks at the failure: 1. Mitigation — add --exit to the mocha command in src/package.json. Mocha's default (--exit=false) waits for the event loop to drain after tests complete. The hard kill happens during that drain or the inter-spec transition. With --exit, mocha calls process.exit(failures) directly once the run finishes, closing the cleanup-race window. Linux/Windows-22/25 are green today, so the natural-drain path is not surfacing real leaks worth preserving. Verified locally: `cd src && pnpm test` -> 1121 passing, 0 failing, 23s. 2. Capture — set NODE_OPTIONS in each Backend tests step to --report-on-fatalerror, --report-uncaught-exception, --report-on-signal, --report-compact, plus --report-directory=${{ github.workspace }}/node-report. If Node crashes at the C++ level (segfault, V8 abort, libuv panic) the runtime writes a JSON diagnostic report with the V8 stack, libuv handle table, JS heap state, and OS info. A new "Upload Node diagnostic reports on failure" step (actions/upload-artifact@v7, `if: failure()`, `if-no-files-found: ignore`) uploads that directory as an artifact per matrix cell — the data we have been unable to capture from JS instrumentation alone. If (1) eliminates the flake on the next push to develop, great. If not, (2) finally gives us the crash dump and we can fix the root cause. Touches all four Backend tests jobs (Linux × 2, Windows × 2). The Windows steps now also set `shell: bash` so the same `mkdir -p ...` line works under git-bash; the existing `working-directory: src` is preserved. NODE_OPTIONS is scoped to the test step only, so the existing vitest step is unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * test(ci): address Qodo review on PR #7748 Three findings from Qodo's review of the Windows + Node 24 flake fix: 1. (bug) "mocha --exit masks handle leaks". --exit removes the post-suite event-loop drain that surfaces leaked timers / sockets — exactly the class of regression that lowerCasePadIds.ts notes is otherwise visible. Linux/local runs are currently green on natural drain, so dropping that signal everywhere would silence real leaks to fix a Windows-only flake. Fix: scope --exit to Windows only. - Remove --exit from src/package.json's "test" script (shared with local dev + Linux CI; both keep natural-drain behaviour). - Append `pnpm test -- --exit` in just the two Windows backend-test steps so the mitigation only runs where the flake actually lives. 2. (observability) "diagnostics exit-matrix misleading". With --exit on Windows, "only exit fires" becomes the EXPECTED pattern there, not a sign of unexpected process.exit(). Update the matrix comment in tests/backend/diagnostics.ts to spell out: clean drain on Linux/local → beforeExit + exit; Windows under --exit → only exit; "only exit" elsewhere still implies an unexpected process.exit somewhere. 3. (rule violation) "no regression test for the mitigation". Repo convention (see admin-i18n-source-lint.test.ts) is to pin policy-bearing config with a source-lint spec so a future refactor can't silently revert it. Add src/tests/backend-new/specs/backend-tests-flake-mitigation.test.ts: - Asserts every "Run the backend tests" step sets NODE_OPTIONS with the report-on-fatalerror diag flags AND is followed by an Upload Node diagnostic reports step (4 of each in the current matrix). - Asserts exactly 2 Windows jobs invoke `pnpm test -- --exit`. - Asserts the shared mocha "test" script in src/package.json does NOT bake in --exit globally. Verified locally: - cd src && pnpm exec vitest run tests/backend-new/specs/backend-tests-flake-mitigation.test.ts → 3 passed (3). - Stream.ts spec without --exit → 31 passing, diag prints "beforeExit code=0" + "exit code=0" (clean drain restored). - Existing admin-i18n-source-lint suite still passes. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| ISSUE_TEMPLATE | ||
| workflows | ||
| dependabot.yml | ||
| FUNDING.yml | ||
| PULL_REQUEST_TEMPLATE.md | ||